Natural language web search API for AI agents, with Jev intent detection and reranking
Send a request in plain language. TypeSafe’s Jev model works out the intent, picks the engines and time window, then reranks every result by relevance before it reaches your agent.
Endpoint
One natural-language request in, a reranked list of sources out. TypeSafe’s Jev model handles intent detection, engine routing, and relevance scoring.
Keyword query and time window from the request.
Up to 5 of 12 engines, searched in parallel.
Every result scored; off-topic results dropped.
Useful for
LLM agent tools / RAG retrieval / Community research
Intent detection, engine routing, and reranking in one call
With the Search API your code chooses the engine and keywords. Ask moves those decisions into the API. A request such as "what are developers saying about Bun 1.3 this month" is typically routed to Hacker News, Reddit, and X plus general web engines, with a one-month window; "recent papers on speculative decoding" goes to arXiv. Results from every engine are scored against the request, so your LLM reads sources that are on topic instead of whatever ranked first on each engine.
AI Intent Detection
TypeSafe’s Jev model reads the request and picks the keyword query and time window. Pass time_range to set the window yourself.
Smart Engine Routing
Jev scores 12 engines, including Google, Reddit, Hacker News, GitHub, arXiv, X, and YouTube, and searches up to five in parallel. Pass sources to choose them yourself.
AI Reranking
Jev scores every result for relevance. Results below 0.5 are dropped, duplicates are merged, and each result keeps its score.
Flat 5 Credits, Model Included
A completed request costs 5 credits however many engines it searches. If no engine completes, the request is not charged.
What is TypeSafe Jev, and why does Ask use it?
Jev is TypeSafe AI’s System One model. Instead of generating text, it answers typed questions with calibrated probabilities, which makes it a good fit for search decisions. For every request, Ask puts typed questions to Jev: which candidate keyword query best matches what the user means (and, for catalogues such as IMDb, the exact title), how well each of the 12 engines fits the request, and whether the request implies a time window. Those answers drive intent detection and engine routing. After the search, Jev is asked whether each returned result is about the subject; that probability is the relevance score used to rerank and filter results. Because Jev returns decisions rather than prose, Ask never writes an answer of its own. Ask is a Search1API product that calls Jev through TypeSafe’s API, falling back to the same model on Cloudflare Workers AI during outages; it is not an official TypeSafe product.
Ask pricing
Ask charges a flat 5 credits per completed request, whether it searches one engine or five, and the Jev model calls are included. If some engines fail while others complete, the failures are listed in errors and the request is charged. If no engine completes, the API returns 502 and nothing is charged. New accounts start with 100 free credits, enough for 20 Ask requests.
Flat 5 credits per request, model calls included.
No charge when no engine completes (HTTP 502).
100 free credits on signup: 20 Ask requests, no credit card.
API key required; up to 30 requests per minute per account.
How an Ask request runs
Send a request in plain language. TypeSafe’s Jev model works out the intent, picks the engines and time window, then reranks every result by relevance before it reaches your agent.
Intent detection: Jev picks the keyword query, scores each engine for fit, and decides on a time window. sources and time_range override these choices.
Parallel search: up to five engines, from Google and DuckDuckGo to Reddit, GitHub, arXiv, and YouTube, are searched at once.
AI reranking: Jev scores every result, results below 0.5 are dropped, duplicates are merged, and the list is sorted by relevance.
Best for
Natural language search API: TypeSafe Jev detects intent, routes to the right engines, and reranks every result by relevance
LLM agents that need one search tool: pass the user’s question through and get reranked, citable sources back.
RAG retrieval where choosing the right sources matters more than collecting the most results.
Research and monitoring across Reddit, Hacker News, X, GitHub, and arXiv, where the right sources change with every question.
FAQ
What is a natural language search API?
A natural language search API accepts a request written the way a person would ask it and decides how to search for it. Ask turns the request into a keyword query, chooses the engines and time window, and returns results reranked by how well they match the request.
How does Ask detect search intent?
Ask asks TypeSafe’s Jev model typed questions about the request: which keyword query fits best, how likely each engine is to help, and whether a time window is implied. Engines are chosen by those probabilities, up to five. The intent field in the response shows the keywords, engines, and window that were used.
How does Ask rerank results?
Jev scores each result from 0 to 1 for whether it is about the subject of the request. Results below 0.5 are dropped, duplicates across engines are merged, and the rest are sorted by score. Each result keeps its relevance score so you can apply a stricter cut-off. The score is a model judgment, not a fact check.
When should I use Ask instead of Search?
Use Ask when you know what you want but not where to look. Use Search when you already know the engine and keywords and want that engine’s own ranking, filters, pagination, or Deep Search content.
How many credits does Ask API use?
A completed Ask request costs a flat 5 credits, however many engines it searches, with the Jev model calls included. If no engine completes, the API returns 502 and the request is not charged.
Does Ask generate an answer?
No. Ask returns reranked results with titles, URLs, snippets, and relevance scores so your own LLM can read and cite them. It does not write a summary or an answer.
What data does Ask send to the Jev model?
Your request and the titles and snippets of the returned results are sent to the Jev model for intent detection and scoring, through TypeSafe’s API or, as a fallback, Cloudflare Workers AI. The search itself runs through Search1API.
Can I use Ask from the SDKs or MCP?
Not yet. Call POST /ask over HTTP with an API key. The official SDKs, CLI, and MCP server do not include Ask, and it is not available through pay-per-request payments.